Integration of Deep Learning in Baseline PET/CT and Multi-Omics Data for Prognostic Modeling in Diffuse Large B-Cell Lymphoma: the Clinicalpet Lymphplex Model

[18F]-Fluorodeoxyglucose (FDG)-PET/CT is essential for staging and evaluating treatment response in diffuse large B-cell lymphoma (DLBCL). The prognostic potential of PET-derived biomarkers requires efficient extraction and analysis of biological signatures. In this study, 18F-FDG-PET scans of 839 newly diagnosed DLBCL patients were analyzed, with DNA sequencing performed on 710 patients. Using the nnUNet deep learning framework, trained on both AutoPET public and in-house datasets, enabled precise PET scan segmentation and biomarker extraction. Key biomarkers, including total metabolic tumor volume (TMTV), Max MTV (MAX_MTV), standardized maximum tumor dissemination (SDmax_patient), and standardized maximum distance between the largest and another lesion (SDmax_bulk), showed significant prognostic value. Integrating LymphPlex genetic subtypes with PET biomarkers and clinical risk factors, we identified critical prognostic indicators: EZB-like-MYC+, MCD-like, TP53Mut subtypes, high TMTV, and elevated lactate dehydrogenase (LDH). Consequently, the ClinicalPET LymphPlex model was developed, effectively differentiating survival rates across various treatments. Additionally, combined with transcriptomic data, we revealed that risk factors within the ClinicalPET LymphPlex like high TMTV and LDH were notably associated with immune-suppressive tumor microenvironments.